Google Gemini's new agent-based video analysis cuts token usage by up to 88 percent
decoder ·
Google is adding agent-based video analysis to its Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite models. Instead of sampling video at a fixed rate, the system autonomously selects relevant sections and decides whether to inspect frames, audio, or transcripts. Google says the approach can detect subsecond events, locate scenes in hours of footage, identify anomalies, and count repeated movements or objects more precisely. On the 1H-VideoQA and LVBench benchmarks, agent-based analysis reduced token usage by up to 88 percent while slightly improving accuracy, and Google reported lower costs on its evaluations. The feature is available for video uploads and YouTube videos through the Gemini API, with developers able to enable it using an “agentic” processing mode at standard token rates. Google plans to add the capability to the Gemini app and use it for the upcoming Ask YouTube feature.
decoder